math/big: regression in BitLen

[rsc]

CL 450055 replaced big.Int.BitLen with a portable constant-time operation for fear of bits.LeadingZeros using a lookup table on some architectures (notably not on x86, which uses a hardware instruction). @drchase's benchmarks found https://perf.golang.org/search?q=upload%3A20221130.13+pkg%3Agithub.com%2Fericlagergren%2Fdecimal%2Fbenchmarks, non-crypto code with a significant performance hit for a property (constant-time) it didn't need.

Talked to @FiloSottile, and he is going to send a CL removing use of big.Int.BitLen from crypto code and restoring the old implementation.

Filing issue so we remember to submit the CL before the release.

[ericlagergren]

Sorry :)

[gopherbot]

Change https://go.dev/cl/454617 mentions this issue: math/big: fix BitLen performance regression

[prattmic]

Fix confirmed on https://perf.golang.org/dashboard/?benchmark=Pi%2Ffoo%3Dericlagergren_%28Go%29%2Fprec%3D100-8&days=60&end=2022-12-21T16%3A46